Evaluation protocol
Methodology
A result is meaningful when another team can repeat the run. Each track defines the input, the held-out reference and the conditions for comparison.
Neural dynamics
Forecasting neural activity
A microisland is a small culture of interconnected neurons. In the F2F task, the model observes a prefix of their activity and predicts the spike times in its continuation.
Keep the forecast in the future
Choose the observation cutoff before evaluation. Fit preprocessing and select model settings using training data. Keep every sample used to make a prediction on the observed side of the cutoff, including the context used to derive features.
Match events neuron by neuron
Match predictions one-to-one with held-out spikes within a declared timing tolerance. A matched event is a true positive. Unmatched predictions are false positives, and unmatched reference spikes are false negatives.
Include precision, recall, the forecast horizon, timing tolerance and averaging rule. State how empty event windows are handled. Report results on future windows from a known circuit and results on unseen circuits separately.
Explore the F2F taskConnectomics
Reconstructing dense tissue
A segmentation pipeline maps an image volume to neuron instances. Evaluate the complete reconstruction against a proofread reference in the same coordinate system.
Time the complete pipeline
Start timing when the pipeline begins reading its input. Stop after it writes the final instance labels. Count each evaluated input voxel once, including when processing uses overlapping tiles. Report the device model, device count, memory, storage and cache conditions.
Measure continuity along neurons
Expected run length (ERL) asks how far an error-free reconstruction extends along the reference skeleton containing a randomly chosen point of cable. Longer correct runs contribute more. Splits shorten runs, and merges invalidate affected segments.
Report ERL in micrometres together with throughput in millions of voxels per second. Two pipelines can trade speed for continuity. Compare both on the same volume and hardware profile.
Explore the segmentation taskReporting a run
- Identify the method
- Give the model or pipeline name, version, configuration and evaluation split.
- Describe the conditions
- Record the allowed inputs, scoring settings and compute environment needed to reproduce the run.
- Verify the result
- Each leaderboard groups comparable runs within its track. A verified entry pairs the reported metrics with the conditions used to obtain them.
Cite UploadBench
Eon Systems PBC. UploadBench. 2026. https://benchmark.eon.systems/